The AI winter is very much over, and we're back to the good old days of selling the future. I bet this team is very sharp, but there's still merit to "over-promise, under-deliver."
"Phoenix, the co-founder, says Vicarious aims beyond image recognition. He said the next milestone will be creating a computer that can understand not just shapes and objects, but the textures associated with them. For example, a computer might understand “chair.” It might also comprehend “ice.” Vicarious wants to create a computer that will understand a request like “show me a chair made of ice.”
Phoenix hopes that, eventually, Vicarious’s computers will learn to how to cure diseases, create cheap, renewable energy, and perform the jobs that employ most human beings. “We tell investors that right now, human beings are doing a lot of things that computers should be able to do,” he says."
Funny to think that instead of curing diseases or making cheap renewable energy, we'd instead try to spend resources to invent a computer to do that for us...
We do that for ourselves when we learn first about something, plan ahead, and then do it. Thing is, we don't know how to cure diseases, we don't know how to make cheap renewable energy, and we certainly don't know how to turn Earth into Heaven. The direct approach might very well be harder than the AI one.
Why do you think our creators made the simulated universe we live in? There's an infinite-loop bug, however: each simulation tries to solve the real problem by creating a sub-simulation.
That's not infinite loop bug, that's recursion. Usually an exit condition will get out of the loop. In this case, sounds like once found solution to cure disease and new energy, the condition is met. Yes?
So the question you should be asking is...if everyone agrees that cheap renewable energy and curing diseases is a good thing, why haven't we done it yet? I guess if you're cynical, you'd argue, because no one can make a profit from it (uh...right).
However, if you see it as a pragmatic problem, then maybe the answer to why not? is because we need better ways to process information -- and this kind of unsupervised machine learning is critical to doing that.
I've given some thought to this, and the answer is it might be. Consider a future where we've managed to eliminate all diseases. By then our natural resistance to disease might have atrophied to a point where we will absolutely be unprepared to deal with new diseases (biologically), before we've managed to develop a cure for them.
Well it is incredibly dangerous, but if we did make safe, super intelligent AI, it would definitely be a much greater investment than directly working on those problems.
To put it lightly, it would automate and massively bring down the cost/increase the speed of research and engineering. Sure enough money, spent on enough humans, given enough time, could eventually find a cure for any disease. But why do that when you can just ask the AI and have a cure overnight?
Of course it's not clear how safe such an AI would be (such a being could easily outsmart us and get what it wants, whatever that even is), nor how difficult it will be to create one.
Well, let's hope they will go broke wasting money into the AI research black hole before being able to actually build one :) Otherwise, you are correct. A successful built of super intelligent AI is probably an extinction event for the humans.
Realistically though, a super intelligent AI, would require far more fundamental breakthroughs. People who say otherwise, should take a look at problems in Control theory.
The effect of scaling things up should be interesting though. Perhaps there is a "Phase transition" like thing, where suddenly something awesome happens. This also means that, sadly, Universities would no longer be able to provide adequate resources for research.
What makes you say that? It may seem that it would take a lot of work, but really, how can we know? Often times difficult problems seem obvious in retrospect.
I wish instead of making computers do things that Humans are doing, people put more effort into making the Humans' job, instructing these stupid beasts (I mean the computers), easier.
When Zuckerberg and Ashton Kutcher invest in things, it doesn't really grab my attention. But when Musk does, it really does sound promising.
The company's tech sounds really awesome, being able to perceive texture from photos and interpret objects from it would be so useful in so many real world applications.
In an interview he emphasized that he was a tiny initial investor / had no idea what they were doing (I'm guessing Peter Thiel convinced him to since Stripe was trying to help make their initial vision for Paypal a reality)
I read/markered "On Intelligence" on my train commute to work and have scribbled a bunch of notes in the book. Pretty interesting and I like the basic idea of the memory-prediction framework, invariant representations, "melodies of patterns", focus on neocortex and the whole same general algorithm for all senses.
I haven't had the time to research how far the general idea has gone or if it is relevant at all but the scetched examples were pretty interesting.
I also found the random remark of "consciousness = what it feels like to have a neocortex" interesting.
Glad to see that some smart money is bet in this general direction.
>I also found the random remark of "consciousness = what it feels like to have a neocortex" interesting.
So there's a way it feels to not have a neocortex? Doesn't feeling anything imply you're conscious, which means you don't need a neocortex to be conscious?
An insect "feels" pain. It doesn't however feel retrospective pain, and it doesn't feel the past in the same way we do.
Having a neocortex is like having a 6th sense.
Our taste, smell, hearing, sight, and feeling neurons are all indirectly fed into our brain. With a neocortex, that input is also fed in. It's our "consciousness" feeling.
Thanks, that makes sense. I was thinking of a different definition of "consciousness", the "hard problem" definition. [0] For an ant to feel pain, it would have to be conscious in the hard problem sense.
I can only speculate, but as you may recall, Numenta abandoned their original, belief-propagation-based design, replacing it with a new one based on sparse distributed memory. Dileep had done a lot of work on the original design, and I recall reading that's what Vicarious is using, having licensed it from Numenta. So I think you can put it down to a difference in technical direction between Dileep and Jeff. As far as I know, the split was amicable.
Their marketing and fund-raising approaches also seem to be completely different. I think this was good for both of them, especially if Dileep is licensing from Numenta.
It seems like the two algorithms are very similar, but that RCN's represent information in a more continuous way. Maybe this allows for more flexibility? Do the sizes of the hierarchical / recursive chunks get changed over time now, or is it not a strict hierarchy anymore? Stuff is weird.
"Phoenix hopes that, eventually, Vicarious’s computers will learn to how to cure diseases, create cheap, renewable energy, and perform the jobs that employ most human beings."
This line made me laugh. Which of the three goals is the most likely and desired outcome? (I'll give you a hint, it isnt curing diseases or finding energy)
That's like saying: 'my robots will cure cancer, bring world peace and replace most manual human jobs.'
Agricultural technology already replaced most manual human jobs (from the era when most human jobs were agricultural). Humans found other things to do, and we found ways to use the surplus.
If it gets to the point where there isn't any unskilled labor left to do, we can always choose as a society to vastly expand the welfare state and divvy up at least part of the accumulated surplus to everyone. We have already moved in this direction a bit, and I expect to see more things along the lines of guaranteed minimum income in the future.
"vastly expand the welfare state and divvy up at least part of the accumulated surplus to everyone"
That's a technology we haven't had much luck with so far. We've had economies where "distribution" was related to direct wealth creation (make a sandwich and eat it myself), property & labour (you make a sandwich with my bread,we eat half each), thievery (gimme your sandwich). We've done sharing in small groups, that may have been the paleo-economy. We've doe bits of charity welfarism, redistribution and centralization but never really succeeded at making those work well at a large scale, especially for those supposed to be protected by it.
We disagree about how well redistribution can and has worked at scale. My ultra-compressed (lossy) take is "pretty inefficient, but somewhat effective improving the lives of the less-wealthy".
One of the most dramatic and promising examples of redistribution, GiveDirectly, is actually doing some followup research on the effectiveness of their redistribution, and it looks pretty good so far: (pdf warning) http://web.mit.edu/joha/www/publications/haushofer_shapiro_u...
That's an extreme example - a relatively-small scale transfer from wealthy donors to a much poorer country - but it speaks well to the principle.
A - Most of those attempts have had mixed results to put it mildly.
B - Scale in economics is a big deal. Obviously you can transfer wealth from one person to another pretty effectively but what happens to an economy, government, society, etc when it's the main income source is a different kettle of fish.
I didn't say impossible. But it's a technology we need to make a big leap on. Money itself is a technology. Maybe we need money itself to be disrupted to overcome some apparent limitations.
I think there are pretty strong correlations between smaller social/political units and more effective, efficient, and non-corrupt welfare/redistribution systems.
There isn't a Nordic country with a population greater than just the population of the New York metro area (let alone New York state...let alone the United States as a whole)
Even in Nordic countries "normal" is working for a living. They have big social and governmental institutions that have a lot of money passing through them and they manage to do that relatively efficiently. But, they don't have a complete disconnect between wealth creation by normal means (owning productive property and/or working) and consuming that wealth. The government is just more involved in the process.
If most people work, pay taxes and use the "free" public transport you still have a situation where most people are both funding the transportation and using it. Consumers & producers of stuff.
These futuristic ideas about AI doing all the work while most people are unnecessary creates a completely different jar of pickled fish.
As in an Intelligence Explosion (http://intelligence.org/ie-faq/.) "Success" is really a bad way of wording it, I just mean if it happens, none of those things will matter. Regardless whether it is friendly or not. Either we go extinct or the AI is so far beyond us, our present progress doesn't make much difference.
If they do pull that off, I hope they will be very, very careful. You know, Intelligence explosion, Friendly AI, taking over the world, that sort of things.
"Information has been running on a primate platform, evolving according to its own agenda. In a sense, we have a symbiotic relationship to a non-material being which we call language. We think it's ours, and we think we control it. It's time-sharing a primate nervous system, and evolving towards its own conclusions."
I'm not sure what your link has to do with your quote… Anyway, this blog post is not quite right.
While I agree capitalism is a more pressing problem than AI right now, it won't kill us all in 5 minutes. A self-improving AI… we won't even see it coming. There is also much more brainpower dedicated to "fixing" industrial capitalism, than addressing existential risks such as AI. And industrial capitalism doesn't need fixing, it needs to be abolished altogether.
Corporations are even less autonomous than the author thinks. Sure, kill a CEO, and some other shark will take its place. On the other hand, those sharks are all from the same families. Power is still hereditary.
If the people were truly informed about how the current system works, it would collapse in minutes. To take only one example, Fractional Reserve Banking is such fraud that if everyone suddenly knew about it, there would be some serious "civil unrest", to put it mildly.
The same does not apply to an AI. It's just too powerful. Picture the how much smarter we are from chimps. Now take an army of chimps, and a small tribe of cavemen (and women), which somehow want to exterminate each other. Well, the chimps don't stand a chance, if the humans have any time to prepare. They have fire, sticks, lances… Their telepathy have unmatched accuracy (you know, speech). And they can predict the future far better than the chimps. Now picture how much more intelligent than us an AI would be.
It's way worse.
---
Now, this new agey speak about information taking a life of its own… It doesn't work that way. Sure, there's an optimization process at work and it is not any particular human brain. But this optimization process is nowhere near as dangerous as a fully recursive one (that is, an optimization process that optimizes itself). And for that to happen, we need to crack a few mathematical hurdles first, like Löb's theorem.
But that's not the hard part. The hard part is to figure out what goals we should program into the AI. Not only we need to pin them down to mathematical precision, but we don't even know what humanity wants. We don't even know what "what humanity wants" even mean. Hell, we don't even know if it has any meaning at all. Well, we're not completely blind, we have intuitions, and a relatively common sense of morality. But there's still a long road ahead.
The connection between the Hostile AI link and the McKenna quote is this: the informational barrier between humans, institutions and technology is highly permeable, and creates a perfect petri dish for natural selection in informational life (you can model them as "memes", although the analogy to genes isn't a perfect one).
Yes, it breeds far less rapidly than a Kurzweilian AGI, and one day we will face that music for better or worse. But what I'm driving at this is that will not come as a singular moment when SkyNet gets the switch flipped; it will be a gradual evolution from the pre-existing emergent intelligence of the "human+institution+technology" informational network. (Even if you had a day where you flipped the switch on an infinitely accelerating AI, that life form would still inherit the legacy data of humans and their institutions, which would inevitably shape its consciousness, infecting it with any memes sticky enough to cross the barrier.)
Hopefully they'll be nice enough to relieve us of these meat costumes and allow us to ascend to their level where civilization and machine can merge into one conscious entity and we can float around the sun for a few million years charging our batteries while we calculate a path through space that leads to our longest survival. When you can simulate the multiverse is there really any reason to travel through space to look for other life? Perhaps it may be interesting to come into contact with another super-consciousness drifting through space, but even then, would they really have much to offer? We would have simulated every occurrence of that too. The only thing left to do would be to somehow transcend space and time which I think is probably impossible.
Have we cracked the brain's "programming language" yet? I am affraid that until now research has been focused on the biological side of it; and it makes more sense to me to replicate the logic instead of replicating the brain's biological structure.
I believe that dataflow/reactive programming is the answer and the direction to follow as its principles are pretty close to how neurons work; plus it can be made to work on top of von neuman architectures.
This comes under the category of neuromorphic engineering. It is an excellent question, of which I am trying to find an answer for months now!
I bet there is more, just buried under varied publications, and I am sure that they created a DSL for their specialized brain based chip.
The most obvious and battle tested way to program a NeuroSynaptic hardware is to create models of the brain (any application) in an algorithm and burning it with an HDL into an fpga or an fpaa. For computation of numerical entities, a small controller running a customized embedded software is used.
By replicating the biological structure, they might shed some light on the logic. Right now we're making almost zero traction, a serious effort to copy it might at least make it clear what we're trying to understand.
At the very least, if they have a cortex in software, it would create a ethical (?) way to experiment with the logic by enabling/disabling pieces and seeing what happens.
Its true, but there are even fields of science based almost entirely on assumptions.
Even then, there is a neuroscience branch called neural coding that apparently acknowledges the existence of a neural code; but judging from the wikipedia entry their approach seems still too "low level".
"Replicating the neocortex, the part of the brain that sees, controls the body, understands language and does math. Translate the neocortex into computer code and 'you have a computer that thinks like a person,” says Vicarious co-founder Scott Phoenix.'"
Do you? Other than the language part it sounds like you may instead have an electronic lizard or cow. Add language and you might have an electronic parrot or dolphin(they can do some language processing).
Something's missing - the ability to reason: deduction, induction and abduction. The ability to set goals and to find a path to those goals. These are the magic that everyone has been seeking and not finding for a long time.
The pieces the Vicarious found speaks of are available today. We have exquisite computer vision, pretty good language understanding and fair robots but no strong AI and certainly no embodied AI. The promises above are hollow. But it will make some people a lot of money.
Don't those other animals not even have neocortices? Or if they do, they are very thin and smooth? I was under the impression that the human neocortex is abnormally thick and wrinkled, making it plausible that the reasoning capabilities are in fact in there.
The neocortex doesn't directly control the body or see, but there are bits that light up when we do. Lizards have the other parts of their brain, directly wired in. We have those parts, too, in addition to our neocortex.
Well it's not clear where such high level functions come from, but it's certainly progress. Making an AI as smart as an animal would be an incredible advancement btw.
Of course it would be. Human brains are only slightly different than animal brains. Most of the work goes into getting to chimps, then it's just a short distance to humans. Animal brains have incredible pattern recognition and reinforcement learning. We don't have to take the path evolution did of course, but it would be progress.
First time I've ever heard of Mark Zuckerberg investing in anything, separate from Facebook. He always cited 'focus' as the reason why he never does it.
Call it what it is, an expert system, market research, a database of decisions/observations. Real "artificial intelligence" only exists in science fiction, in the minds of children playing with toys. Your computer (doll) won't ever love you back or have any awareness or understanding no matter how bad you want it to. It's a cool sounding buzzword for marketing, but if there's any intelligence here it's coming from a few developers hiding behind their tricky algorithms.
A computer will never have intelligence, no matter how many factors and randomizations you code in to give the illusion of intelligence. Calling a collection of observations "intelligence" is an insult and severe underestimation of what intelligence is. If you believe artificial intelligence is possible, you're missing out on what life has to offer--or you would never think a box of switches could come alive.
There's no hint of evidence you = your brain. It's safe to say the brain processes information literally. But we have no idea where intelligence originates. Sadly, some people never get beyond a literal interpretation of things.
> But we have no idea where intelligence originates.
It seems to me we don't even have a good definition of intelligence, never mind an understanding of it. You admit it yourself, so why the diatribe on how (not even why) "real artificial intelligence" is impossible? You haven't defined what AI is nor demonstrated why it will never happen.
Man will never fly. You can't prove that birds fly because of their wings or air pressure. Therefore I am 100% certain that it's magic and can never be achieved by mere machinery.
Calling a air pressure differential machine "flying" is an insult and severe underestimation of what flying is. If you believe artificial flying is possible, you're missing out on what life has to offer--or you would never think a box of gears could come alive.
Sadly, some people confuse their ignorance with knowledge and make all kinds of embarrassing claims.
>There's no hint of evidence you = your brain.
Sign up for a lobotomy we'll see much "you" is left afterwords.
> If you believe artificial intelligence is possible, you're missing out on what life has to offer--or you would never think a box of switches could come alive.
You got me. Believing x would make me think life has less meaning. Therefore, x is false. What an argument.
Correlation does not imply causation. If I unplug your LCD, that doesn't imply the application failed. If I pull out your CPU, doesn't imply the cloud app is not still functioning. Possible examples of this, people who are "brain dead" that have reached out to grab a scalpel (organs about to be harvested) or have come back to life after brain death or pronounced dead but can quote a conversation that happened in the room while brain dead, etc.
This idea is absurd which I believe arises from a certain type of perspective that humanity needs to go beyond. Humanity needs to be less caged in perception. Let me expound on a point of view that maybe different to what see.
To separate matter, and mind is a paradoxical argument, because they're both of the same thing. Going back to the old idea of the fallen tree, if there's no mind then matter does not exist, and if matter doesn't exist then mind can not arise.
To put in other terms, if there's nothing receiving the projection, then what is the projection projecting on? Projection, and reception are another way of looking at mind, and matter. Mind being reception, matter being projection.
So going by that logic, and assuming that we're all made of matter, we can say that matter itself is both projection, and reception. So if matter is both projection, and reception, then what does that mean? Are we all "just" matter? Yes. Exactly.
But the argument isn't whether or not we're made of matter. I think we all agree that we're made of matter. I think the argument is that we humans share a certain inexorable feeling of qualia that arises from being human. Yes that's it. It's that qualia that distinguishes us from the rest of everything, except...
The problem is that qualia arises from our material form. Of course assuming that everything is matter, and the idea of the eternal soul, or other such argument, is false. Then that means qualia itself is matter.
Ok. What the hell am I getting at?
Maybe matter is more complex, more interesting than we perceive. Maybe matter itself is "intelligent", and it's just another form distinct from human perception. hmmm... So am I saying that everything that is matter is "intelligent"? Yes. That's exactly what I'm saying, BUT there are different forms of material patterns that form different constructs intelligence.
Meaning that how we receive, or in what form we receive the projection determines our perspective. Right now it just so happens we humans have a POV of humans.
The thing is that due to our incredible ability to not just receive, but to also project what we receive onto different things gives us the power of empathy. The illusion that we can perceive from a different POV. That we can somehow distill our perspective, and project it onto another thing. It's worked quite well so far. Mathematics, language, science, etc. But once we try to see from another perspective that's unimaginably different then it all breaks down.
Let's try to look at the perspective of ant for instance. Well we can't, because if you think about it you can't think of non-thought. Think of non-thinking, is an oxymoron. An ant doesn't think, I mean I'm sure it thinks, but it has completely different sense organs, a completely different set of logical processes, it has a completely different structure, and a completely different perspective than humans. It's unimaginable, because we can only view it from our perspective, which in its renders the idea false. We can only view the world from our perspective. Yet we can't call the ant unintelligent, an ant is very intelligent.
What we see is just that, and what we see differently, is still just seeing. We can't stop seeing, and once we stop seeing, then we stop being human. A human being is just another form of seeing, ants another, computers yet another. Everything has intelligence, it's just not in a recognizable form. In a relatable form. We're all just a box of switches. A mesh of material patterns that filters through existence to produce being. Demeaning different forms of being as lesser is a very human centric perspective. See differently, from the top of the mount, and realize you'll only ever see like a human being.
> Computers based on neuromorphic design are the best bet for intelligent Machines.
I wouldn't go that far. We don't understand enough about the nature of intelligence and the way brain works; right now saying that "the best bet for AI is for computer to look like a brain" is like saying "the best bet for heavier-than-air flight is for a machine to flap wings like birds", which was a stupid idea for the reasons we now understand well.
Neuromorphic computers do not look like a brain. They just borrow some of it's so called 'features'.
I am not saying that we should copy the brain. But at least we could copy the design, just like we did for aeroplanes. Neuromorphic sensors could act like our cerebellum, which act during unforeseen incidents. They are typically error tolerating.
I wonder if you could make it completely analog. Find functions that can be done fast/cheaply in silicon, and then design learning algorithms that can take advantage of them.
* First Law: A robot must never harm a human being or, through inaction, allow any human to come to harm.
* Second Law: A robot must obey the orders given to them by human beings, except where such orders violate the First Law.
* Third Law: A robot must protect its own existence unless this violates the First or Second Laws.
* Zeroth Law: A robot must never harm humanity or, by inaction allow humanity to come to harm.
Every other law gets an unless this interferes with the zeroth law. suffix.
I encourage anyone to read the robots series, specifically( in that order ): The Caves of Steel, The Naked Sun
and The Robots of Dawn, where the three laws are used in the story, and even the zeroth law is implied in the third book.
There are a lot of problems with these laws. The main problem is that such an AI would be completely dominated by the first law. It would spend all it's time and resources in order to even slightly decrease the probability of a human coming to harm. Nothing else would matter in it's decision process since the first law has priority.
Second, how would you implement such laws in a real AI, especially the type they are building? This requires defining increasingly abstract concepts perfectly (what is "harm"? What is "human being"? What configuration of atoms is that? How do you define "configuration of atoms"? Etc.) And this is pretty much impossible to begin with in reinforcement learners, which is what is currently being worked on. All a reinforcement learner does is what it believes will get it a reward or avoid pain/punishment. Such an AI would just steal the reward button and run away with it, or try to manipulate you to press it more. It doesn't care about abstract goals like that.
You do realize that, as stated, these laws are (1) practically impossible to implement, (2) routinely broken by humans (especially the first law - life-sacing and cosmetic surgery, piercings, sport, euthanasia, abortion), and (3) a matter of philosophical/moral subject, decisions about which, IMO, should be in the ___domain of humans, not robots.
He explored many issues, e.g. what happens when robots misinterpret the laws, or what should very expensive robots do, or what happens if robots interpret emotional pain as "harm", but I'm not sure he investigated the obvious, yet extremely hard issue of encoding the laws from human language into computer program.
In the books the robots adhered to the laws strictly. The problem was that humans were able to circumvent the laws rather easily. For example; lie to the robot or divide the murder trough many robots each unaware of each other.
In absence of humans the robots were perfect for deciding moral subjects( as long they have enough information ), the opposite what tomp is suggesting.
Humans "operate" using emotions and logical biases, but computers "operate" using logic. To implement the first law, you must be certain that there is always something that an agent can do or must not do in order to "save" humans. This is almost always not true (hence moral disagreements).
Also, even if you change the laws to get rid of logical inconsistencies, you still have to translate the words into logic, by strictly defining them, which is again impossible (as humans disagree what these words mean).
Elon musk read Asimov's in his childhood. I hope he will stand for those values. And also, I think such powerful projects should be open sourced for the public.
Probably not. If Noam Chomsky is to be believed (I do), most research to date has been publicly funded. In the US, it has been mostly through military expenditures. (To take only one example, ARPANET itself was funded by the military.)
The actually awesome part is having huge investments on long term research. Private or public, it doesn't make any difference.
In response to your comment on another topic:
You can run a "freedom box" as follows:
http://freedombone.uk.to
The guide will work for a raspberry pi or a beagle bone.
It was created out of frustration with progress of the freedom box project.
I know someone who intereviewed at Vicarious and came away unimpressed. That said, any company with an investment by a guy who can make his company buy it out is a good one to invest in.
It's like if you said, "I like ice cream" and I reported it as, "This person thinks buying ice cream is more noble than giving a starving family a bag of rice."
Look, guys, sure, in some sense computing is part of
the best promise for AI. Fine. I'll even agree that
at least for now computing is necessary.
But, note, nearly everything we've done in computing,
especially in Silicon Valley for the past 15 years, has
been to apply routine software development to
work that we already well understood how to do
manually. A small fraction of the efforts have been
some excursions into more, but these have been
relatively few and with rarely very impressive
results. Net, what Silicon Valley does know how to
do is build, say, SnapChat (right, it keeps the NSA
spooks busy looking at the SnapChat intercepts
from Sweden!).
But for anything that should be called AI,
there is another challenge that is very much
necessary -- how to do that. Or, if you will,
write the software design documents from the
top down to the level of the individual
programming statements. Problem is, very likely
and apparently, no one knows how the heck to do
that.
Given a candidate design, people should want to
review it, and about the only way to convince
people, e.g., short of the running software passing
the Turing test or some such, is to write out the
design in terms of mathematics. Basically the only
solid approach is via mathematics; essentially
everything else is heuristics to be validated only
in practice, that is, an implementation and not
a design.
Thing is, I very much doubt that anyone knows how
to write a design with such mathematics. If so,
then long ago there should have been such in
an AI journal or with DARPA funding.
Basically, bluntly, no one knows how to write
software for anything real about AI. Sorry 'bout that.
Wby? We just do not know hardly anything about
how the brain works. We don't know more about how
the human brain works than my kitty cat
knows about how my computer works. Sorry 'bout that. And
AI software will have a heck of a time catching up
with my kitty cat.
By analogy, we don't know more about how to
program AI than Leonardo da Vinci knew about
how to build a Boeing 777. Heck the Russians
didn't even know how to build an SR-71. Da Vinci
could draw a picture of a flying machine, but
he had no clue about how to build one. Heck,
Langley fell into the Potomac River! Instead, the
Wright brothers built a useful wind tunnel (didn't
understand Reynolds number), actually were able to
calculate lift, drag, thrust, and engine horsepower,
and had found a solution to three axis control --
Langley failed at those challenges, and da Vinci
was lost much farther back in the woods.
We now know how our daughters can avoid
cervical cancer. Before the solution, "we dance
'round and 'round and suppose, and the secret
sits in the middle, and knows.", and we didn't
know. Well, the cause was HPV, and now there
is a vaccine. Progress. Possible? Yes. Easy?
No. AI? We're not close enough to be in the
same solar system. F'get about AI.
Well we do actually have a purely mathematical approach to AI worked out. Granted it requires an infinite computer, and personally I don't think it will lead to practical algorithms. But still, it exists. And from the practical side of things, machine learning is making progress in leaps and bounds. As is our understanding of the brain.
Remember that airplanes weren't built by Da Vinci because he didn't have engines to power them. It wasn't that long after engines were invented that we got airplanes. The equivalent for AI, computing power, is already here or at least getting pretty close.
> Well we do actually have a purely mathematical approach to AI worked out.
Supposedly with enough computer power and
enough data, a one stroke solution to
everything is stochastic optimal control,
but that solution takes, say, just brute
force to, say, planetary motion instead of
Newton's second law of motion and law of
gravity. Else, need to insert such laws
into the software, but we would insert only laws humans
knew from the past, or have the AI software
discover such laws, not so promising. This
stochastic optimal control approach is
not practical or even very insightful.
But it is mathematical.
> machine learning is making progress in leaps and bounds.
I looked at Prof Ng's machine learning course, and
all I saw was some old intermediate statistics,
in particular, maximum likelihood estimation (MLE),
done badly. I doubt that we have any
solid foundation to build on for any significantly
new and powerful techniques for machine learning.
I see nothing in machine learning that promises
to be anything like human intelligence. Sure,
we can write a really good chess program, but
no way do we believe that its internals are
anything like human intelligence.
> As is our understanding of the brain.
Right, there are lots of neurons. And if
someone gets a really big injury
just above their left ear, then we have a good
guess at what the more obvious results will be.
But that's not much understanding of
how the brain actually works.
It's a little like we have a car,
have no idea what's under the
hood, and are asked to build a car. Maybe
we are good with metal working, but
we don't even know what a connecting rod is.
> It wasn't that long after engines were invented that we got airplanes.
The rest needed was relatively simple, the wind tunnel,
some spruce wood, glue, linen, paint, wire, and
good carpentry. For the equivalent parts of AI,
I doubt that we have even a weak little hollow
hint of a tiny clue.
In some of the old work in AI, it was said that
a core challenge was the 'representation problem'.
If all that was meant was just what programming
language data structures to use, then that was
not significant progress.
Or, sure, we have a shot at understanding the
'sensors' and 'transducers' that are connected
to the brain: Sensors: Pain, sound, sight,
taste, etc. Transducers: Muscles, speech,
eye focus, etc. We know some about how the
middle and inner ear handles sound and
the gross parts of the eye. And if we show
a guy a picture of a pretty girl, then we can
see what parts of his brain become more
active. And we know that there are neurons
firing. But so far it seems that that's about
it. So, that's like my computer: For sensors
and transducers it has a keyboard, mouse, speakers,
printer, Ethernet connection, etc. And if we
look deep inside then we see a lot of circuits and transistors.
But my kitty cat has no idea at all about the
internals of the software that runs in my computer,
and by analogy I see no understanding of the
analogous details inside a human brain.
Or, we have computers, and we can write software for
them using If-Then-Else, Do-While, Call-Return, etc.,
but for writing software comparable with a human
brain we don't know the first character to type
into an empty file for the software. In simple
terms, we don't have a software design. Or,
it's like we are still in the sixth grade,
have learned, say, Python, and are asked to
write software to solve the ordinary differential
equations of space flight to the outer planets --
we don't know where to start. Or, closer in,
we're asked to write software to solve the
Navier-Stokes equations -- once we get much
past toy problems, our grid software goes
unstable and gives wacko results.
Net, we just don't yet know how to program
anything like real, human intelligence.
I was referring to AIXI as the perfect mathematical AI.
The main recent advancement in machine learning is deep learning. It's advanced the state of the art in machine vision and speech recognition quite a bit. Machine learning is on a spectrum from "statistics" with simple models and low dimensional data, to "AI" with complicated models and high dimensional data.
>if someone gets a really big injury just above their left ear, then we have a good guess at what the more obvious results will be. But that's not much understanding of how the brain actually works.
Neuroscience is a bit beyond that. I believe there are also some large projects like Blue Brain working on the problem.
I swear I saw a video somewhere of a simulation of a neocortex that could do IQ test type questions and respond just like a human. But the point is we do have more than nothing.
I looked it up: His 'decision theory' is essentially
just stochastic optimal control. I've seen
elsewhere claims that stochastic optimal control
is a universal solution to the best possible AI.
Of course, need some probability distributions;
in some cases in practice, have those.
That reference also has
> Solomonoff’s theory of universal induction formally solves the problem of sequence prediction for unknown prior distribution.
Hmm? Then the text says that this solution is
not computable -- sound bad!
Such grand, maybe impossible, things are not
nearly the only way to exploit mathematics
to know more about what the heck we are doing
in AI, etc.
Approximations to AIXI are possible and have actually played pacman pretty well. However I still think solomonoff induction is too inefficient in the real world. But AIXI does bring up a lot of real problems with building any AI, like preference solipsism and the anvil problem, and designing utility functions for it.
> I was referring to AIXI as the perfect mathematical AI.
I will have to Google AIXI. A big point about
being mathematical is that that is about the only
solid way we can evaluate candidate work before running
software and, say, something like a Turing test.
Some math is most of why we know, well before any
software is written, that (1) heap sort will run
in n ln(n), (2) AVL trees find leaves in ln(n),
and (3) our calculations for navigating a space
craft to the outer planets will work. More
generally, the math is 'deductive' in a
severe and powerful sense and, thus, about the
only tool we have to know well in advance of, say,
writing a lot of software.
But math does not have 'truth' and, instead,
needs hypotheses. So, for hypotheses, for some
design for some software for AI, we need some.
Enough hypotheses are going to be a bit tough
to find. And math gives only some mathematical
conclusions, and we will need to know that these
are sufficient for AI; for that we will want,
likely need, a sufficiently clear definition
of AI, that is, something better than just an
empirical test such as a Turing test or
doing well on and IQ test. Tough challenge.
Instead of such usage of math, about all we have
in AI for a 'methodology' is, (1) here I have
some intuitive ideas I like, (2) with a lot
of money I can write the code and, maybe,
get it to run, and (3) trust me, that program
can read a plane geometry book with all the
proofs cut out and, then, fill in all the
proofs or some such. So, steps (1) and (2)
are, in the opinion of anyone else, say, DARPA,
'long shots', and (3)
will be heavily in the eye of the beholder.
The challenges of (1), (2), and (3) already
make AI an unpromising direction.
> The main recent advancement in machine
learning is deep learning.
It's advanced the state of the art
in machine vision and speech recognition quite a bit.
AI has been talking about 'deep knowledge' for a
long time. That was, say, in a program that
could diagnose car problems, 'knowledge' that
the engine connected to the transmission
connected to the drive shaft connected to the
differential connected to the rear wheels or
some such and, then, be able to use this
'knowledge' in 'reasoning' to diagnose
problems. E.g., a vibration could be
caused by worn U-joints.
When I was in AI, when I worked
in the field, there were plenty of people who
saw the importance of such 'deep knowledge'
but had next to nothing on really how to
make it real.
For 'deep learning', the last I heard, that was
tweaking the parameters 'deep' in some big
'neural network', basically a case of nonlinear
curve fitting. Somehow I just don't accept
that such a 'neural network' is nearly all that
makes a human brain work; that is, I'd expect
to see some promising 'organization'
at a higher level than just the little
elements for the nonlinear curve fitting.
E.g., for speech recognition, I believe an important
part of how humans do it is to take what they
heard, which is often quite noisy and by itself
just not nearly enough, and
compare it with what they know about the
subject under discussion and, then, based
on that 'background knowledge', correct the
noisy parts of what they heard. E.g., if
the subject is a cake recipe for a party
for six people, then it's not "a cup of salt"
but maybe a cup or two or three of flour.
If the subject is the history of US
presidents and war, then "I'll be j..."
may be LBJ and "sson" maybe "Nixon". Here
the speech recognition is heavily
from a base of 'subject understanding'.
An issue will be, how the heck does the
human brain sometimes make such
'corrections' so darned fast.
For image recognition, the situation has
to be in part similar but more so: I doubt that
we have even a shot at image recognition
without a 'prior library' of 'object
possibilities': That is, if we are
looking at an image, say, from
satellite, of some woods and looking
for a Russian tank hidden there,
then we need to know what a Russian
tank looks like so that we can guess
what a hidden Russian tank would look
like on the image so that we can, then,
look for that on the image. Here we
have to understand lighting, shadows,
what a Russian tank looks like from
various directions, etc. So, we are
using some real 'human knowledge'
of the real thing, the tank, we
are looking for.
E.g., my kitty cat has a food
tray. He knows well the difference
between that tray and everything
else that might be around it --
jug of detergent, toaster, bottle
of soda pop, decorative vase,
a kitchen timer. Then I can move
his food tray, and he doesn't get
confused at all. Net, what he is
doing with image recognition is
not just simplistic and, instead,
has within it a 'concept' of his
food tray, a concept that he
created. He's not stupid you know!
So, I begin to conclude that for
speech and image recognition, e.g.,
handwriting recognition, we need
a large 'base' of 'prior human
knowledge' about the 'subject area',
e.g., with 'concepts', etc.,
before we start. That is, we need
close to 'full, real AI' just to,
say, do well reading any handwriting.
From such considerations, I believe
we have a very long way to go.
Broadly one of my first cut guesses about
how to proceed would be to roll back
to something simpler in two respects.
First, start with brains smaller, hopefully
simpler, than those of humans.
Maybe start with a worm and work up to
a frog, bird, ..., in a few centuries, a
kitty cat! Second, start with the baby
animal and see how it learns once it
starts to as an egg, once it's born,
what it gets from its mother, etc.
So, eventually work up to software that
could start learning with just "Ma ma'
and proceed from there. But can't just
start with humans and "Ma ma" because
a human just born likely already has
somehow built in a lot that is crucial
we just don't have a clue about.
So, start with worms, frogs, birds,
etc.
Another idea for how to proceed is to
try for just simple 'cognition'
with just text and image input
and just text output. E.g., start
with something that can diagram
English sentences and move from there
to some 'understanding', e.g., have
made progress enough with
'meaning' that, e.g., know
when two sentences with quite different
words and grammar really mean essentially
the same thing and when they don't mean
the same thing report why not and be
able to revise one of the sentences so that
the two do mean the same thing.
So, here we are essentially assuming
that AI has to stand on some capabilities
with language -- do kitty cats have
an 'internal language'? Hmm ...! If
kitty cats don't have such an 'internal
language', then I am totally stuck!
Then with some text and
image input, the thing should be able
to cook up a good proof of the
Pythagorean theorem.
I can believe that
some software can diagram
English sentences or come close
to it, but that is just a tiny
start on what I am suggesting.
The real challenge, as I am guessing,
is to have the software keep track of
and manipulate 'meaning', whatever
the heck that is.
And I would anticipate
a 'bootstrap' approach: Postulate and
program something for doing such things
with meaning, 'teach' it, and then
look at the 'connections'
it has built internally, say, between
words and meaning, and also observe
that the thing appears to work well.
So, it's a 'bootstrap' because it
works without our having any
very good prior idea just why;
that is, we could not prove in
advance that it could work.
So, for kitty cat knowledge,
have it understand its environment
in terms of 'concepts' (part of 'meaning') hard, soft, strong, weak,
hot, cold, and, then, know when
it can use its claws to hold on to
a soft, strong, not too hot or too
cold surface, push out of the way
a hard, weak obstacle, etc.
Maybe some such research direction
could be made to work.
But I'm not holding my breath waiting.
Keep in mind evolution managed to make strong AI, us, through pretty much blind, random mutations, and inefficient selection.
The thing about deep learning is that it's not just nonlinear curve fitting. It learns increasingly high level features and representations of the input. Recurrent neural networks have the power of a Turing machine. And stuff like dropout are really efficient at generalization. My favorite example is word2vec. Creating a representation for every English word. Subtracting "man" from "king" and adding "woman" gives the representation for "queen".
Speech recognition is moving that way. It outputs a probability distribution of possible words, and a good language model can use that to figure out what is most likely. But even a raw deep learning net should eventually learn those relationships. Same with image recognition. I think you'd be surprised at what is currently possible.
> It learns increasingly high level features and representations of the input.
In the words of Darth Vader, impressive. In
my words, astounding. Perhaps beyond belief.
I'm thrilled if what you say is true, but I'm
tempted to offer you a great, once in a life time
deal on a bridge over the East River.
> The future looks bright.
From 'The Music Man', "I'm reticent. Yes, I'm
reticent." Might want to make sure
no one added some funny stuff to the Kool Aid!
On AI, my 'deep learning' had a good 'training
set', the world of 'expert systems'. My first
cut view was that it was 99 44/100% hype and
half of the rest polluted water. What was left
was some somewhat clever software, say, the Forgy
RETE algorithm. My views
after my first cut view was that my first
cut view was quite generous, that expert systems
filled a much need gap in the literature and would
be illuminating if ignited.
So, from
my 'training set' my Bayesian 'prior probability'
is that nearly anything about AI is at least
99 44/100% hype.
That a bunch of neural network nodes can somehow
in effect develop internally just via adjustments
in the 'weights' or whatever 'parameters' it has
just from analysis of a 'training set' images of
a Russian tank (no doubt complete with
skill at 'spacial relations' where it is claimed
that boys are better than girls) instead of somehow
just 'storing' the data on the tank separately
looks like rewiring the Intel processor
when download a new PDF file instead of just
putting the PDF file in storage. But, maybe
putting the 'recognition means' somehow
'with' the storage means is how it is actually done.
The old Darwinian guess I made was that
early on it was darned important to understand
three dimensions and paths through three dimensions.
So, going after a little animal, and it goes behind
a rock. So, there's a lot of advantage to
understanding the rock as a concept and that
can go the other way around the rock and get
the animal. But it seems that the concept of
a rock stays even outside the context of
chasing prey. So, somehow intelligence
works with concepts such a rocks and also
uses that concept for chasing prey,
turning the rock over and looking under it,
knowing that a rock is hard and dense,
etc.
Net, my view is that AI is darned hard,
so hard that MLE, neural nets, decision
theory, etc. are hardly up to the level
of even baby talk. Just my not very well
informed, intuitive, largely out of date
opinion. But, I have a good track record:
I was correct early on that expert systems are
a junk approach to AI.
There is a degree of hype. They are really good at pattern recognition, maybe even superhuman on some problems and with enough training and data. But certainly they can't "think" in a normal sense or are a magical solution to the AI problem. And like everything in AI, once you understand how it actually works, it may not seem as impressive as it did at first.
>instead of somehow
just 'storing' the data on the tank separately
looks like rewiring the Intel processor
when download a new PDF file instead of just
putting the PDF file in storage.
Good analogy, but how would you even do that? One picture of a tank isn't enough to generalize. Is a tank any image colored green? Is it any object painted camouflage? Is it any vehicle that has a tube protruding from it?
In order to learn, you need a lot of examples, and you need to test a lot of different hypotheses about what a tank is. That's a really difficult problem.
The trouble with this kind of artificial intelligence is that I don't think it's possible to think like a human without actually having the experience of being human.
Sure, I think we could aim to build basically a robot toddler that had a sensory/nervous/endocrine system wired up analogously to ours. It would basically be a baby, and would have to go through all the developmental stages that we go through.
But I suspect we'll have a hard time modeling that stuff well enough to create anything more than "adolescent with a severe learning disability". People underestimate just how carefully tuned we are after millions of years of evolution. The notion that we could replicate that without having another million years of in situ testing and iteration seems naive.
And even then, why would we expect the AI to be smarter than a human? There is already quite a lot
of variation in humans. Many people at the ends of the bell curve have extraordinary processing power in ways typical humans don't. But it turns out while those things are useful in some ways, they limit those people in other ways. So it's not like we're not trying out evolved designs, it's that on balance they don't seem to actually function fundamentally better.
One cool thing about the robot is that you could have many bodies having many experiences all feeding into one brain. But I'm not convinced that would actually lead to "smarter". I mean, look at old people. Yes we get smarter as we age. But age also calcifies the mind. All of that data slowly locks you into a prison of past expectations. In the end, it's a blend of naive and experienced people in a society
that maximizes intelligence. And again, it's not like societies haven't been experimenting with that blend. Cultures have evolved to hit the sweet spot. It's not clear that adding 1000 year old intelligences would help.
And anyway, we already have 1000 year old intelligences: books!
You could say that there is benefit to having all of that in one "head" but then you have to organize it! Which experience drives the decisions, the one from 2014 or the one from 3014?
Again, culture evolved explicitly to solve this problem. People write books and the ones that work stick around.
I guess what I'm saying is the evolution of the human being is already here: it's the human race, fed history via culture, connected by the internet, in symbiosis with computers.
The idea that removing the humans from that system would make it smarter makes no sense to me. Nor does the idea of writing programs to do the jobs that humans do well. It's like creating a basketball team with 5 Shaquille O'Neils. I don't think they'd actually be able to beat a good, diverse team with one or two Shaqs.
Or think of it this way: if numerical/logical aptitude is such a huge advantage in advancing capital-U Understanding, why do smart people bother learning to paint? Why do we bother listening to children? Why do we bother having dogs?
I would argue it's because intelligence is as multi-media as the universe is. Sometimes a PhD has something to learn from a basset hound. And the human race has just as good a handle on it as any AI ever will. We just have a different view of the stage. We have the front row and they have the balcony.
"Phoenix, the co-founder, says Vicarious aims beyond image recognition. He said the next milestone will be creating a computer that can understand not just shapes and objects, but the textures associated with them. For example, a computer might understand “chair.” It might also comprehend “ice.” Vicarious wants to create a computer that will understand a request like “show me a chair made of ice.”
Phoenix hopes that, eventually, Vicarious’s computers will learn to how to cure diseases, create cheap, renewable energy, and perform the jobs that employ most human beings. “We tell investors that right now, human beings are doing a lot of things that computers should be able to do,” he says."